[Paper Review] OrchideaSOL: a dataset of extended instrumental techniques for computer-aided orchestration

Carmine-Emanuele Cella, Daniele Ghisi+4 more|arXiv (Cornell University)|Jul 1, 2020

Music and Audio Processing9 references11 citations

TL;DR

OrchideaSOL is a free, high-fidelity dataset of extended instrumental techniques derived from the original Studio On Line (SOL) recordings, reprocessed to restore accurate dynamic ranges and ensure consistency for target-based computer-aided orchestration. It improves upon prior versions by using the proximity microphone (track 3), resampling to 44.1 kHz, and applying loudness-weighted RMS normalization, enabling more realistic orchestral mixtures in systems like Orchidea.

ABSTRACT

This paper introduces OrchideaSOL, a free dataset of samples of extended instrumental playing techniques, designed to be used as default dataset for the Orchidea framework for target-based computer-aided orchestration. OrchideaSOL is a reduced and modified subset of Studio On Line, or SOL for short, a dataset developed at Ircam between 1996 and 1998. We motivate the reasons behind OrchideaSOL and describe the differences between the original SOL and our dataset. We will also show the work done in improving the dynamic ranges of orchestral families and other aspects of the data.

Motivation & Objective

• To address the degradation of dynamic range consistency in the original SOL dataset, which compromised orchestration results in systems like Orchidea.
• To create a free, reliable, and reproducible dataset for computer-aided orchestration that preserves the true dynamic characteristics of orchestral instruments.
• To improve the fidelity and usability of extended playing techniques for research in music information retrieval, music cognition, and sound synthesis.
• To provide a version of SOL that is freely accessible for non-commercial use, with verifiable integrity via checksum validation and version control.
• To enable more realistic orchestral sound mixtures by correcting amplitude normalization and resampling artifacts present in earlier distributions.

Proposed method

• Reconstructed OrchideaSOL from the original SOL recordings by selecting only the proximity microphone (track 3), which provides the cleanest, least reverberant signal.
• Resampled all audio files from 48 kHz to 44.1 kHz with 24-bit depth to ensure compatibility with standard audio processing pipelines.
• Converted all files to monophonic to eliminate stereo imbalance and reduce data complexity while preserving core timbral and dynamic content.
• Applied loudness-weighted RMS normalization using the pyloudnorm library to correct amplitude inconsistencies across dynamic markings (pp, mf, ff).
• Fitted quadratic regression models to loudness-weighted RMS data to characterize dynamic response across MIDI pitch and dynamic marking, enabling predictive modeling of instrument dynamics.
• Generated and stored pre-computed features including spectral envelopes, MFCCs, spectral peaks, and spectral moments (centroid, spread, skewness, kurtosis) for each sample using 4096-sample windows with 2048-sample hops.

Experimental results

Research questions

• RQ1 How can the dynamic range inconsistencies in the original SOL dataset be corrected to improve the realism of computer-aided orchestration outputs?
• RQ2 To what extent does using a single, consistent microphone channel (proximity mic) improve the reliability and reproducibility of orchestral sound datasets?
• RQ3 Can a free, open, and verifiable version of SOL be created that maintains audio quality while enabling broader research access?
• RQ4 How accurately can the dynamic response of orchestral instruments be modeled using loudness-weighted RMS regression across pitch and dynamic marking?
• RQ5 What is the impact of pre-computed spectral and perceptual features (e.g., MFCCs, spectral moments) on the performance of music information retrieval and generative modeling tasks?

Key findings

• OrchideaSOL contains 12,253 high-quality audio samples representing 35% of the total playing techniques in the original SOL dataset, with extended techniques and ordinario sounds preserved.
• The dataset corrects the amplitude normalization issues of prior versions by using the proximity microphone (track 3), resulting in more accurate dynamic ranges across instruments.
• Loudness-weighted RMS regression models were successfully fitted for each instrument, with the flute’s model showing a quadratic fit: $ L_{\textrm{dB},\textrm{flute}} \approx -82.3301 + 0.2732M + 3.0831D + 0.0194MD + 0.0008M^2 + 0.0823D^2 $.
• Spectral features including average spectra, MFCCs (first 20), spectral envelopes (1024 bins), and spectral peaks (first 120) were pre-computed and made available for all samples.
• The dataset is distributed under a free non-commercial license with verifiable integrity via MD5 checksums through the mirdata library, ensuring reproducibility and trustworthiness.
• OrchideaSOL is hosted on Zenodo with a permanent DOI, enabling citable access and long-term availability for scientific research.